[2H6-OS-2c] OS-2

6:45 PM - 7:00 PM JST(9:45 AM - 10:00 AM UTC)

[2H6-OS-2c-06] Dynamic Objective Function Selection Based on Market States for Financial Decision-Making

〇Keigo Sakurai¹, Takahiro Ogawa¹, Miki Haseyama¹, Anjyu Anan², Kei Nakagawa³ (1. Hokkaido University, 2. Nomura Asset Management Co,Ltd., 3. Osaka Metropolitan University)

Keywords:

Finance,Decision-Making,Objective Function

In financial decision-making tasks such as stock recommendation and portfolio allocation, optimizing an appropriate objective function under estimated future risk and return is critical; however, due to market non-stationarity, a single fixed objective is unlikely to remain optimal across different regimes. Many existing machine learning approaches adopt a fixed objective throughout the investment horizon, while regime-switching methods suffer from estimation noise, delayed switching, and potential instability caused by frequent changes. To address these issues, we propose DOSS (Dynamic Objective Selection with Safeguards), a framework that dynamically selects the objective function at each time step based on observable market statistical features without explicitly inferring latent regimes. DOSS formulates objective selection as a multi-class classification problem over a predefined candidate set and incorporates a confidence-aware gating mechanism that falls back to a conservative default objective when prediction confidence is low. Experiments on public benchmarks demonstrate that DOSS consistently outperforms static-objective methods and LLM-based direct selection approaches in both performance and operational stability.